CONTINUATION AND REFLECTION OF SOLUTIONS OF

PARTIAL DIFFERENTIAL EQUATIONS 1

FRITZ JOHN

A solution of an ordinary differential equation can be continued as

long as its graph stays in the domain, in which the equation is regu-
lar. On the other hand a solution of a partial differential equation can
have a natural boundary interior to the domain of regularity of the
equation. Let R be a closed region and S a portion of the boundary
of R. Then S is a natural boundary for a solution u defined in R, if
there exists no solution defined in a full neighbourhood of a point of
S which agrees with u in R. Examples for the occurrence of such
natural boundaries are well known from the theory of harmonic func-
tions. Neither the equation nor the solution has to show any very
singular behavior on approaching a natural boundary S.
Only in very special situations can one prove t h a t every sufficiently
regular solution known in a region R can be continued across a portion
5 of the boundary. This is e.g. the case for solutions of a single differ-
ential equation which is hyperbolic with respect to 5. It is also the
case for solutions of certain overdetermined systems of equations, like
those associated with analytic functions of several complex variables ;
other systems with this property have been studied by S. Bochner

In more general cases the only solutions for which one can prove
continuibility are those satisfying suitable boundary conditions on 5.
The classical example is furnished by solutions u(x, y) of the Laplace
equation
(la) uxx + uvy = 0
defined for y ^ 0 and satisfying the boundary condition
(lb) «(*, 0) = 0.
They can always be continued across y = 0 by the formula
(lc) u(x, -y) = - u(x, y).

An address delivered before the New York meeting of the Society on April 21,
1956, by invitation of the Committee to Select Hour Speakers for Eastern Sectional
Meetings; received by the editors January 11, 1957.
1
A summary of the results of this paper appeared in Colloques internationaux du
centre national de la recherche scientifique, vol. 71, Nancy, 1956.
327
328 FRITZ JOHN [November

In recent years one has succeeded in extending this continuation theo-

rem in various ways, involving either more general boundary condi-
tions or more general differential equations. I mention the following
generalizations. Thus a sufficiently regular solution of (la) defined for
y^O can be continued across y = 0 whenever it satisfies instead of
(la) a boundary condition of the form
(Id) uy = A(x, u, v, ux)
on ;y = 0, where A is an analytic function of its arguments and v de-
notes the conjugate harmonic to u (See H. Lewy [2], R. Gerber [3]).
A solution of any linear elliptic equation of any order and in any num-
ber of dimensions which has vanishing Dirichlet data on a hyperplane
can be continued across that plane, provided the equation has ana-
lytic coefficients. (See Morrey [5], Nirenberg and Morrey [4].) A
solution of the equation of minimal surfaces that is analytic on an
analytic boundary curve can be continued across that curve. (See
Lewy [15].)
Continuation of solutions constitutes an important tool in finding
an explicit solution of many problems of applied mathematics. It has
been used by Helmholtz, Kirchhoff, Shiftman and others in determin-
ing motions of liquids possessing a free boundary. 2 In these problems
the continuation of the flow is not explicit as long as the location of
the free boundary is not known; the free boundary and the flow are
obtained simultaneously by also taking into account the boundary
conditions on fixed boundaries and reducing the problem to one in
conformai mapping. Explicit local continuation can be used, when
the boundary conditions are sufficiently simple. This situation arises
e.g. in the linearized theory of gravity waves, where one is led to
consider solutions of the Laplace equation (la), which satisfy a
boundary condition of the form
(2) Uy + kU = 0
with constant k on the #-axis.8 The continuation of solutions of
more general second order equations with constant coefficients which
satisfy boundary conditions of the type (2) on a plane has been
studied by Diaz and Ludford [9].
A few cases in which solutions of higher order equations can be con-
tinued by an explicit reflection formula have been established. Porit-
sky [lO] proved that a solution u(x, y) of the bi-harmonic equation
(3a) A2u = 0
2
See Lamb [6], Ch. IV, Milne-Thomson [7].
8
See Stoker [8].
1957] SOLUTIONS OF PARTIAL DIFFERENTIAL EQUATIONS 329

which satisfies the boundary conditions

(3b) u = uy = 0 for y = 0
can be continued across the x-axis using the formula
(3c) u(x, —y) = - u(x, y) + 2yuy(x, y) — y2Au(x, y).
Analogous formulae have been obtained by Duffin [ l l ; 12] for the
systems of equations describing static equilibrium of a 3-dimensional
isotropic elastic body. Finally Huber [13] has generalized formula
(3c) to the case of solutions of a poly-harmonic equation
(4) àPu{xx • • • , * » ) = ()
having vanishing Dirichlet data on a hyper-plane.
It seems desirable to have a more systematic study of the types of
boundary conditions that guarantee existence of a continuation of the
solution of a differential equation and also of the cases in which such
continuation can be achieved by an explicit formula. Of course, such
formulae can be expected only for very simple situations, say for
linear equations with constant coefficients and for solutions satisfying
linear homogeneous boundary conditions with constant coefficients
on a plane boundary. I should like to present here some results in this
direction for a rather simplified situation: The equation shall be
linear and homogeneous and have constant coefficients, the boundary
shall be plane, and the boundary conditions shall be of the type that
the solution and a certain number of its derivatives vanish on the
boundary plane. The results have been extended by R. I. Canavan
to the most general equation with constant coefficients in his thesis
on Necessary conditions for continuation and reflection principles for
solutions of linear partial differential equations with constant coefficients,
New York University, 1957.
There shall be w + 1 independent variables XQ, X\, > Xfi» We
write x for the n- vector (#i, • • - , # n ), and U(XQ, X) for the function
u(xoy xi, - • • , xn). The differential equation shall be of the form
(5a) P(U &u(xo, x) = 0
where P(£o> £) is a form with constant complex coefficients of degree
Nm
£o = d/dxo and £ = (d/dxi • • • , d/dXn).
Moreover we make the assumption
(5b) P ( l , 0) = 1,
330 FRITZ JOHN [November

equivalent to noncharacteristic character of the plane x0 = 0. The

boundary condition shall be of the form
k
(5c) (£ow)s0=o = 0 for k = 0, • • • , s — 1.
We shall show that in general if s is too small there exist solutions of
(Sa, c) which are in Coo for Xo §; 0 and have the plane Xo = 0 as natural
boundary.
Solutions of partial differential equations with natural boundaries
can be obtained in the form of series similar to those used for con-
structing analytic functions with natural boundaries. We consider the
characteristic equation of equation (5a) :
(6) P(vo, V) = 0.
For given rj = (rji, • • • , rjn) the N roots rj0 of (6) may be denoted by
Xi, • • - , Xiv. Assume now t h a t for a certain real rj with 77 • rj = 1 the
roots 770 of (6) satisfy the conditions
(7a) Im Xi è lm X2, • • • , Im \8+i ^ 0
and
(7b) Im Xi > 0.
Let f(x0) be the solution of the ordinary differential equation
«4-1
n «o - A*)/(*O) = 0
*-i

with initial conditions

* (0 for k = 0, • - • , s - 1,
£o/l*o) = < _ .
U for k =s
for #o = 0. Then, because of (7a), f(xo) and any of its derivatives be-
come infinite at most like a power of Xo for xo—•+ 00. The function
eir>xf(xo)

is a solution of (5a, c). It follows t h a t

00

(8) u(xo, *) = 2 j *~;'2 e x P (i2ifi-x)f(2*x0)

for Xo^O is again a solution of (Sa, c) and is of class C*. We shall

prove that this function u has xo = 0 as natural boundary.
Since 775^0 there exists a real vector y such that
i957] SOLUTIONS OF PARTIAL DIFFERENTIAL EQUATIONS 331

y-ri = 2T.

We have for any positive integer r and for x0 ^ 0 from (8)

(9) U(XQ, x + 2~ry) — U(XQ, X) = wr(ffo, #),
where
r-l
(9a) wr(*o, *) = 12 e~3'2 e x P (î2^-*)/(2^o)(exp(i2'- r + 1 ir) - 1)
y-o

is an entire analytic solution of (5a).

If u(xo, x) would not have the plane #o = 0 as natural boundary
there would exist a point x = X, a positive number e, and a function
U(xo, x) of class CN in
(10) |* - X\ < €, |*o| <e
which satisfies (5a) and agrees with u(x0f x) for x0^0.
The expression
V(x0, x) = U(xo, x + 2~ry) — U(xo, x) — wr(xo, x)
is a solution of (5a) defined for
| x - X | < e/2, | *o | < €
provided r is so large that
2'>2|y|/€.
Moreover F(#o, #) = 0 for # 0 è 0 . It follows from the uniqueness theo-
rem of Holmgren 4 that there exists a positive e', independent of r,
such t h a t F(x 0 , #) = 0 for all x0, x with
| « - X\ < e', |*0| <€'.

In other words, the functional equation (9) satisfied by u must be

preserved under continuation.
Let (?(£o, £) denote the differential operator
«+i

Since

I I (to - *A*)/(*o) = *iXl*°,

fc«=2

4
See John [14].
332 FRITZ JOHN [November

we have for | # - - X | < € / 2 , — e < # o < 0

| Qtto, &[U(xo, x + 2~y) - U(x0, x)] \
r-l

X) 2*3'e->2 exp(i2>x-rj)(exp (i2^ r+1 7r) - 1) exp (i2>\ix0)

^ 2 . 2 - ^ ° exp ( - ( r - l ) 2 - 2r-1x0 Im (Xi))

00

- 2 exp (-2-- 2 xo lm (X0) Z 28,'e-'2.

y-o
Hence
lim | £)(£<,, Ö[f(*o, x + 2~ry) - U(x0, x)]\ = w.
T—*cO

But this is incompatible with U(x0, x) belonging to class CV. Conse-

quently u(xo, x) has the plane x0 = 0 as natural boundary, if we can
find a real rj with | -171 = 1 such that the corresponding roots X* of the
characteristic equation (6) satisfy (7a, b). Since equation (6) is homo-
geneous in rjo and 77, the restriction 1771 = 1 is unessential.
We call boundary conditions for a solution on the plane #0 = 0
adequate for continuation, if every sufficiently regular solution defined
on one side of the plane in a neighborhood of the plane can be con-
tinued locally across that plane. We can then formulate our result as
follows :
For the boundary conditions (5c) to be adequate for continuation of a
solution of (5a) it is necessary that for every real rj the roots X& are either
all realy or that at least N—s of them have positive real part and at least
N — s of them have negative real part.
Consequently for s<N/2, the boundary conditions (5c) can be
adequate for continuation only if for every real rj the roots X& are
all real, or, in other words, they are never adequate unless the differ-
ential equation (5a) is hyperbolic with respect to the plane x0 = 0.
If N is even and 5 = N/2, conditions (5c) express that u has vanish-
ing Dirichlet data on #o = 0. Then the vanishing of the Dirichlet data
can be adequate for continuation only if for any real rj the solutions
770 of (6) are either all real or all imaginary. They are all imaginary
when equation (5a) is elliptic. In this case vanishing of the Dirichlet
data is adequate for continuation by virtue of the result of Morrey
and Nirenberg mentioned above. It is possible that the necessary
condition for adequacy of boundary data is always sufficient, regard-
less of the type of the equation.
We apply the condition to the example of the equation of elastic
waves in an istropic medium:
19571 SOLUTIONS OF PARTIAL DIFFERENTIAL EQUATIONS 333

(11a) (^ - C*A) ( | i - C " A ) «(*, y, « , 0 - 0

where A is the Laplace operator in xyz-space. Here the boundary

conditions
du
(lib) u = — = 0 for # = 0
a*
are inadequate for continuation, if the constants c and c' are distinct.
For, in that case, there exist real y, z, t such that the 4 roots of the
characteristic equation
{t2 - c\x2 + y2 + z2)){t2 - c'2(x2 + y2 + z2)) = 0
are neither all real nor all imaginary. If, on the other hand, c = c', the
conditions ( l i b ) are adequate; indeed, then, continuation is achieved
by the explicit formula (analogous to (3c))
x2 /d2 \
(lie) u(—x, y, z, t) = — u + 2xux + — ( c2A)u
c2\dt2 )
where on the right hand side the arguments of u are x, y, z, t.
Equation (11a) is hyperbolic with respect to the plane £ = 0. Thus
the boundary conditions
(lid) u = ut = 0 for / = 0
are adequate for continuation, even when CT^C'. Here again an explicit
"reflection formula" exists, giving the continuation, which will be
derived later (see formula (32c)).
We next turn to the question : Can continuation be achieved by a
simple formula of the type (lc) or (3c), if the data are adequate for
continuation? Instead of trying to give directly a more precise mean-
ing to what constitutes a "simple formula" it is convenient to take
notice of a more qualitative feature of formulae (lc), (3c), namely
that they constitute "reflection principles." The values of the con-
tinued function at one point P are expressible in terms of the values
of the original functions and of its derivatives in points on the line
through P perpendicular to the boundary plane. We shall restrict
ourselves to the case, where the order N of the differential equation
is even, iV=2m, and where the boundary conditions on the plane
Xo = 0 consist in the vanishing of the Dirichlet data (s = m) :

(12) (£o«)»o-o = 0 for k = 0, • • • , m - 1.

We are led to the following definition: Equation (5a) is said to
334 FRITZ JOHN [November

possess a reflection principle (with respect to the Dirichlet data on

the plane Xo = 0), if there exists a number e > 0 (depending only on
P) such that every solution u(x0, x) of (Sa), which is analytic in a
cylindrical set which has the domain D in #-space as base
(13a) x in Z>, 0 g xo < h

can be continued as a regular analytic solution into the set

(13b) xin D, -eh < x0 ^ 0,
provided u satisfies the boundary conditions (12) for x in D.
It turns out that all equations having a reflection principle in this
sense can be characterized rather simply, and that existence of a
reflection principle implies existence of an explicit formula for con-
tinuation of a solution.
It follows from the definition that an equation (5a) possesses no
reflection principle, if there exists an analytic solution u(xo, x) of
(5a) with vanishing Dirichlet data on x 0 = 0, which is regular for
x = 0, x 0 ^ 0 , and has a singularity in some point of the negative x0-
axis. Let now rj be any complex w-vector?^0, and let Xi, • • • , X2m be
the corresponding roots rj0 of (6). Then

XQ
u(xo, x) =
H (ri'X+ XkX0 + Xi)
*=i

(_l)m+l y- dX

2iri •
*s o
/ . -
_.
(rj-x + \Xo + Xi) 1 1 (X — Xjfc)
fc=4

(where C is a small path about the point X= — (rj-x+\i)/xo) is easily

seen to be an analytic solution of (5a), (12) which is singular at the
point
x = 0, xo = — 1.

If this solution is regular for x = 0, and #o^O, there is no reflection

principle. This is the case, when
X&#o + Xi 9e 0 for real ^ ^ 0 and k = 1, • • • , m + 1.
Thus necessary for the existence of a reflection principle is that when-
ever one root Xi is different from 0 then at least one out of any m of
i957l SOLUTIONS OF PARTIAL DIFFERENTIAL EQUATIONS 335

the remaining 2 m — 1 roots lies on the ray opposite to that containing

Xi. In other words, it is necessary that for any complex rj either the
roots Xjfc of (6) all vanish, or that none of them vanishes and that they
all lie on the same line through the origin, half of them on either side
of the origin.
We are led to look for the forms P(rjo, v) whose roots T7O=XA; lie on
a straight line through the origin for any complex rj. There would
have to exist a real 6 and real Ck such that
X* = ckeid for k = 1, • • • , 2m.
Let

where qk(ri) is a form of degree k in rj = (rju • • • , r)n) ; qk{t])e~ike would

have to be real for any k = 1, • • • , 2m. Hence for any complex rj and
any k, j=l, • • • , 2m

(14) ^ ^ -
(&0l))*
would have to be real, whenever gy(ry) 3^0. It follows that the expres-
sion (14) must be constant, since an analytic function of rj which is
real for all complex rj in the neighbourhood of a point in 77-space must
be a constant.
Let 5 be the subset of the set of integers j= 1, • • • , 2m, for which
q^rj) 7^0. Unless P = rjlm, the set S is not empty. Let ju be the greatest
common divisor of all the numbers in 5. We can represent JJL in the
form

with integral coefficients n3: Let Q = Q(r]) be the rational function

defined by
Q = I I SiGl)"'.
yes
Then for any k in S
Çk(v)» TT gifa)'" f f , .
— 11 — constant = (7*)**
QW Tes &(,)*•"
or
(is) «bh) = TfcOoew^
M
where (Y&) is independent of 77 and F^O. Since the polynomial ^(^7)
336 FRITZ JOHN [November

is bounded for all bounded rj, the same holds for the rational function
Q(rj), which consequently must be a polynomial in TJ. The number
k/n is an integer for k in S, and thus, by (15), 7* is independent of rj.
It follows that P is of the form

(16) i> = Z vT^CrQiv)'.

Here the cr are constant, and Q is a form of degree fx. Whenever

770=X is a root of the polynomial P given by (16), cX is also a root,
where e is any /xth root of unity. For XT^O, M > 2 , the various eX do
not all lie on a straight line. Hence P is either given by rff* or by (16),
where Q is a form of degree /x with /x = 1 or /x = 2. In either case we can
factor P , and conclude that a reflection principle is possible only if
P is of one of the following 3 types:

2m

(i7b) p = n o»o - a w ) ,
m

(17c) i> = I I (no - «*?(*)).

Jfc=l

Here the a* are constant, and L and g denote, respectively, a linear

or a quadratic form in rj which is independent of k. Since the roots
rjo must either vanish or lie on one line through the origin, half on
each side, we can change L or q by a constant factor, so that in (17b)
the ah are all real, ai, • • • , am positive, a m+ i, • • • , a2m negative, and
in (17c) so that all the a* are real and positive.
It remains to show that for any equation of one of the types
(17a, b, c) with real ah of the correct sign there actually exists a re-
flection principle. This is achieved by constructing an explicit formula
that represents a continuation of the desired type. We omit the case
of an equation of type (17a) in which continuation is trivial from
further consideration. To arrive a t reflection formulae for equations
of type (17b, c), we first derive such a formula for the case of an ordi-
nary differential equation with constant coefficients.
Let p(\) be a polynomial of degree 2m with roots ai, • • • , #2™,
where a*5^ay for k = l, • • • , m and j = tn + l, • • • , 2m, and all the
ai are different from 0. We consider a solution u(x0) of the equation
(18a) P(i;o)u(xo) = 0
for which
(18b) (£o«(*o))*o-o = 0 for k = 0, 1, • • - , m - 1.
19571 SOLUTIONS OF PARTIAL DIFFERENTIAL EQUATIONS 337

We put
m 2m
(18c) A(\) = I I (* - «»), B(\) = £ (X - A»),
&=1 fc=m+l

so that £(X) =i4(X)5(X). Let C and C" be two closed paths in the X-
plane not containing the origin which are exterior to each other, and
such that C' contains #i, • • • , am and C" contains a m+ i, • • • , aim.
We define for any X, JJL the polynomial RÇK, JJL) by

X — [x
Then for any integer j ^ 0 the identity

£o«(#o) = (2iri)~2d) da(p d/3

J c J e
(19)

is satisfied for every solution u of (18a, b). The expression on the

righthand side of (19) is clearly defined since any solution u of (18a)
is analytic, and (R(a//3)!-o, ex) is a differential operator depending regu-
larly on the parameters a, j8.
Identity (19) can be verified most easily by observing that the
general solution u of (18a, b) is of the form
h(\)e*xo
(20) u(xQ) = — : <h "v%'\ d\,
2-iriJc
2irij c p(\)
where h(K) is an arbitrary polynomial in X of degree <ra, and C is a
path containing all roots of p. We may assume that C contains both
C and C" in its interior. Then
(a \ ( P \ 1 f p(X) - p(a)
\P*' ) \a ) 2wiJc (X-a)*(X)
p(a) f h(\)
= h(a)eP*» - ^-^-A — eP^«d\.
2iriJc (X - a)p(\)
The second term makes no contribution, since
P(a)
çpXxo/a
(fi - a)A(a)(\ - a)
33S FRITZ JOHN [November

is a regular function of a for a inside C', j8 on C", X on C and the

point a = 0 outside C'. Moreover, from the theory of residues
1 C Ha) j h{0)
IriJc 0 8 - a) A (a) A(fi)
since h is of lower degree than A, and a = / 3 is the only pole of the
integrand outside C'. It follows that

ftR\— £o,aW — *o)

Evaluating the remaining portion of the right hand side of (19), in

a similar way, the validity of the identity (19) becomes obvious if
we compare it with the expression for &U(XQ) arising from (20).
In principle the right hand side of identity (19) can be evaluated
by the theory of residues. It then yields a representation of %QU{XQ) in
terms of u and its derivatives taken for the arguments a,kX0/aj and
a3'Xo/ak where k = l, • • • , m and j = m + l, • • • , 2m. If all ak are
positive all aj negative, this represents a reflection principle, since all
ük/aj are then negative and since for x0<0 we can express ^0u(xo) in
terms of u and its derivatives for positive arguments. If all a*, ay are
distinct we have, for example, the formula
m 2m
£ou(xo) = 2 iC
(20a)
y „/<**& \ /arx0\ jn/<h£o \ ( ak \
arRI f ak IuI I — akRI ; ar)u[ — Xo)
\ ar / \ ak / \ ak / \ar /
(ar — ak)A'(ak)Br(ar)
Here
a
n/ tèo \ (orxA r p{d/ds) ~]
R[ > ak)u[ 1= — u(s)
\ ar / \ ak / La/as — ak J s»ars0/a*
If we differentiate identity (19) k times with respect to Xo and put
Xo = 0 we obtain the relation

(21) [£o+**(*o)k=o = [ F * ( £ ) «(*)]

where Fy*(X) is a polynomial of degree ^2m — l-\-k:
1957] SOLUTIONS OF PARTIAL DIFFERENTIAL EQUATIONS 339

(22) Fjk(X) = (2Ti)-2(f da(f d$-

J c J e 08 ~ a)A(a)B(p)
There exist polynomials Gjk(X) and HjkÇK) such that
(23) X>+* = Fjk(\) + Gjk(\) + Hjk(\)p(\)
where Gjk is of degree <2m. Then from (21) we have
[Gy*(£o)«(*o)]«o-o = 0
for any solution u of (18a, b). Since for such a u the values of
[£ow(#o)]*o«o vanish for fe = 0, • • • , m — 1 and are arbitrary for
k = rn, m + 1, • • • , 2m — 1 , it follows that Gjk must be of degree <m.
Hence there exists an identity of the form (23) with a polynomial
Gjk of degree < m.
We are now in a position to write down a reflection principle for
an equation of the type (17b). Let u{x^ x) be a solution of the differ-
ential equation
2m

(24) Pfto, Ö«(*o, *) = I I tto - a*L(0)*(*o, *) = 0

which is analytic for real x0, x in a set (13a), and satisfies the Dirichlet
conditions (12) for x in D. Here the d^ shall be real numbers,
#i, • • • , dm positive, am+i, • • • , a2m negative, and L shall be a linear
form with constant (real or complex) coefficients.
Formally the conditions on u are of the type (18a, b), only with ak
replaced by akL{%). For j — 2tn — 1, we are then led to the identity
/ p»\ 2m— 1
(25) £o w(*o, #) = fl(*o, x)
where
(26a) v(x0, x) = (27ri)-2(b dad) dp ;
Jc> Je» Qî-a)A(a)B((l)

P ^ y Éo,*)-P(aI(Ö,Ö
:2»t—1 .
5 = /9 WI XQ, X 1

(26b)
* ((—^ f .o ., **)) -- P(PL{Ç),Ç) , ,
a2m—1 \a / (a \
U X( X
a
\J » )-
340 FRITZ JOHN [November

Here the polynomials A (a), 5(/3) are defined by (18c); the path C
can be chosen close to the smallest interval on the positive real axis
containing the points ai, • • • , am and C" close to the smallest interval
on the negative real axis containing the points am+u • • • , #2™. Then
a/j3 and /3/a are always close to a negative real number whose ab-
solute value does not exceed the largest of the values —#&/#/. Let e
be the positive number defined by
1 a
— = TVT
Max ~ * •
€ *J=l,---,2m dj

Then for a on C", j8 in C" the quantities a//3 and fi/a will be close to
negative real numbers lying in the interval from — 1/e to 0. Further-
more, for x in D and — eh<x0^0 the quantities u((a/(3)x0, x) and
w((j3/a)#o, oc) will be defined as well as 5 which represents a differen-
tial expression on u and depends regularly on the parameters a, (3.
(The corresponding formula for j<2m — 1 would still contain the
factor Z,(£) ;-2m+1 , which would not be a differential operator.) I t
follows that the function v(xo, x) is defined and analytic for real Xo, x
in the set (13b). In order to prove that it represents a continuation of
^lm"1u it is sufficient to compare the derivatives with respect to x 0
at *o = 0. Comparison with formula (22) shows that this amounts to
proving the identity

(27) [èT~1+ku(xo, x)]xo=o = I L2m~1+\è)F2m-i kl —— —Ju(s, x)

for j = 2m — 1, & = 0, 1, 2, • • • and any solution u of (24), (12). By

(23) we have
2m-l+k 2m-l+k -1 2m-l+k -1
Ç0 — ^ r2m-lk\-L SO) -T L (jT2m-lk\-L, Co)

+ Lk~1H2m.1 kiL^^L^piL'^o)
identically in £0 and L. For L = L(%) both sides represent differential
operators. We apply this identity to a solution of (24), (12) obtaining
(27) since
L^iOpiL-'moMxo, x) = Pfto, &u(xo, x) = 0
and since also
[L~- 1(ÖG2m_i k{L-im,)u{x,, x)]Xo=o = 0,
Lm~l(J)G2m~i *(i""1(f)?o) being a polynomial of degree < m . This com-
pletes the proof of the fact that v(x0, x) represents an analytic con-
tinuation of t$n~1u(xQ, x).
1957] SOLUTIONS OF PARTIAL DIFFERENTIAL EQUATIONS 341

The function v is given by a double integral in the complex a- and

j8-planes; the integrand has poles at the zeros of A {a) and B(/3). The
integral can be expressed by the residues at these poles, which in turn
can be written as differential operators acting on the integrand. In
principle then v(x0, x) can be expressed as a combination of deriva-
tives of u taken at the points

— Xo, x ) and ( — xo, x )

with & = 1, • • • , m and j = w + l, • • • , 2m. The resulting formula in-

volves only u and a finite number of its derivatives in a number of
real points, all lying on the same perpendicular to the plane x0 = 0.
In case all the a» are distinct the formula is of the form (20a) :
m 2m g
£o u(xo, x) = X) ]C >
k=l r=m-J-l

2m—1
ar
'GH- P(a*£(Ö,Ö
• u(s, I)

as Is—Orlo/ajfe

p(£,t)-P(OrL®,Z)
2m—1
— ate u(s, Ö

ds
- aM 8=ajcXo/ar

1
The identity (25) gives a continuation v of £jf «. If we define a
function Z7(x0, x) in (13b) by the conditions

(28) %T V = v, (&U)Xo=o = ft5w),«-o for k = 0, , 2w - 2,

Z7 and w will agree with all their derivatives at Xo = 0, and hence Z7 will
represent an analytic continuation of u. The determination of U in
terms of v involves only repeated integration along the line XQ = con-
stant. This establishes the existence of a reflection principle in the
sense defined here for equations of the type (17b). It is easily seen
that the resulting formulae yield a continuation of a solution u of
(24), (12) not only when u is analytic, but whenever u is sufficiently
often differentiable.
Reflection principles for equations of type (17c) are obtained from
the same basic formula (19), taken for j = 2w —2. If p(\) is an even
polynomial, with roots
342 FRITZ JOHN [November

X = ± ak for k = 1, • • • , m,
m
we have B(X) = ( — l) A(—X). Replacing j3 by — j3 in (19) we can
choose the same path of integration on C' for a and /3. We then arrive
a t the formula
2m-2 2(-l)m r
2(-i) r r
r oc^m~2
(29) £o u(xo) = ft <b da(b
Ax A d3 # —
2
(27r^)
{2*iyJc*/ c Jc> (a
. / c (a + (3)A(a)A((3)

Ky*0) ~#w
a 2
—2 Jo — a
2 "(T*)'
P
Here for positive real ai, • • • , am the path C' can be taken as a set
of small circles about the points ai, • • • , am.
Let now u(xo, x) be a function which is analytic in the set (13a),
satisfies the differential equation
m
(30) Pft,, &u = I Ifto- «*<?(!))« = 0

and the Dirichlet conditions (12), and let

2(-i)mwr r
2(-i) r
r ap2m~2s
»(*o, *) = -rz-rr-Q
22 àato dp —
(31) (2m)
W ^J ' c> J
^ c' (a +
• /3)^(a)^(«'

P(^f0,j)-P(a(^)) 1/2 ,0
£=-—^ u ( Xo, x J,
a 2 2

p
then a is defined and analytic in the set (13b), where

€= mm — •

By the same argument as before, comparing derivatives at x 0 = 0, one

verifies that
£o u(x0, x) =fl(#0,x).
Repeated integration with respect to x0 then furnishes an explicit
reflection formula for u(xo, x) itself.
A typical example is given by the case m = 2 with distinct a*. Let
w(#o, #) be a solution of the equation
1957] SOLUTIONS OF PARTIAL DIFFERENTIAL EQUATIONS 343

2 2 2 2
(32a) ({0 — aiff(Ö)(£o — a2q(Ç))u(x0, *) = °>
where q is a quadratic form and #i, 02 are distinct real positive quanti-
ties. Let u satisfy the initial conditions
(32b) u = %QU = 0 for x0 = 0.
The residues at the simple poles of the integrand in the expression
for v are easily evaluated, and one finds the formula
2 2
(fli + a2)(ai — a2) £ou(x0, x)

r / 2 2 d2 2 2 \ 1
= (01 + a 2 ) ( ( a i + a2) — - — 2a1a2q(£) J u(s, x)
(32c) L ^ V *• / J _
— 2aia2 (— »2?(Ö ) «(^, *o)
LW / Js=-a
-a^XQlai

- 2aia2 ( — - aîgtt) ) «(*, *o)

In the case of an equation of the type (32a), for which ai = a2 = l,

one obtains a reflection formula either directly from (31) by evaluat-
ing residues at double poles, or from (32c) by passing to the limit
with ai and a2. The resulting formula is
2 r y d2 a3 2 a4 a
%ou(xo, x) = ( — - + 2x 0 —- — *o—7 + 2g(£) — 4#0?(£) —
Lw2 ds3 ds4 ds
2 d2\ "I

Taking into account the initial conditions (32b) we can obtain a

simpler formula for u itself, which does not involve any quadratures:

K
— 1 — 2x 0
ds
d

XQ 2 h #o<z(£) )M(S, X)
ds /
2d2

J8=-x0
2 \ 1

Formula (33) contains formula (3c) for the bi-harmonic equation as

the special case q= — £?. Reflection formulae for the more general
polyharmonic case can be obtained from (31).
One observes that the reflection formulae for partial differential
equations of the types (17b, c) have been obtained formally from the
formula (19) which applies to ordinary differential equations, by
treating P(£o, £) as an ordinary differential operator in £0. For general
forms P(£o, £) the formal expressions have no immediate concrete
344 FRITZ JOHN

interpretation, since the formal roots £0 of the equation P(£o, £ ) = 0

are only algebraic functions of differential operators. For equations
of the special type (17b, c) meaningful formulae are obtained, due to
the fact t h a t in t h a t case the quotients of roots are ordinary numbers
instead of operators.

BIBLIOGRAPHY
1. S. Bochner, Green's formula and analytic continuation, Ann. of Math. Studies,
vol. 33, 1954, pp. 1-14.
2. H. Lewy, A note on harmonic f unctions and a hydrodynamical application, Proc.
Amer. Math. Soc. vol. 3 (1952) pp. 111-113.
3. R. Gerber, Sur une condition de prolongement analytique des fonctions harmon-
iques, C. R. Acad. Sci. Paris vol. 233 (1955) pp. 1560-1562.
4. L. Nirenberg and C. B. Morrey, On the analyticity of the solutions of linear elliptic
systems of partial differential equations, Technical Report No. 5, Department of Mathe-
matics, University of California, Berkeley, 1956.
5. C. B. Morrey, A variational method in the theory of harmonic integrals, OSR-
Technical Note 55-108, 1955, pp. 18.
6. H. Lamb, Hydrodynamics (reprinted by Dover Publications, 1945).
7. L. M. Milne-Thomson, Theoretical hydrodynamics, 3d éd., Macmillan, 1955.
8. J. J. Stoker, Surface waves in water of variable depth, Quarterly of Applied
Mathematics vol. 5 (1947) pp. 1-54.
9. J. B. Diaz and G. S. S. Ludford, Reflection principles for linear elliptic second
order partial differential equations with constant coefficients, Scritti Matematici offerti
a Mauro Picone, 1955, pp. 87-95.
10. H. Poritsky, Applications of analytic functions to two-dimensional biharmonic
analysis, Trans. Amer. Math. Soc. vol. 59 (1946) pp. 258.
11. R. J. Duffin, Continuation of biharmonic f unctions by reflection. Duke Math. J.
vol. 22 (1955) pp. 313-324.
12. , Analytic continuation in elasticity, Carnegie Institute of Technology,
Technical Report No. 23 CIT-ORD-6D-TR23, 1956.
13. A. Huber, The reflection principle f or polyharmonic functions, Pacific J. Math,
vol. 5 (1955) pp. 433-439.
14. F. John, On linear partial differential equations with analytic coefficients.
Unique continuation of data, Communications on Pure and Applied Mathematics
vol. 2 (1949) pp. 209-253.
15. H. Lewy, On the boundary behavior of minimal surfaces, Proc. Nat. Acad. Sci.
U.S.A. vol. 37 (1951) pp. 103-111.
NEW YORK UNIVERSITY